Claude Fable 5: What It Costs and When to Use It

Claude Fable 5 went generally available on June 9, 2026, and per Anthropic's launch announcement it is the company's most capable generally available model, the first of a new Mythos-class tier that sits above Opus in the lineup rather than replacing it. The API identifier is claude-fable-5, and it ships with a 1M tokens (128k output) window and adaptive thinking always on: the model decides how hard to reason per request, with no extended-thinking toggle to manage.

The practical question is not whether Fable 5 is good, it is. The question is when it earns its price over the rest of the Claude lineup, and how it stacks up against every other current coding model. This guide sticks to verified facts and gives you the honest routing answer.

Data verified July 18, 2026 against Anthropic's published pricing and docs. For scale: $10 in / $50 out per MTok is exactly double Opus 4.8's $5 in / $25 out per MTok, which is why the routing question in the last section matters more than the benchmark bragging rights.

Per Anthropic, Claude Fable 5 and Claude Mythos 5 share the same underlying model. Fable is the generally available version, shipped with additional safety measures around dual-use capabilities. Mythos 5 is invitation-only, reserved for approved organizations in Anthropic's defensive cybersecurity program. For everyone reading this, Fable 5 is the one you can actually use.

No frontier model launch in recent memory has had a first month like this, and it is why so many people are searching for Fable 5 in the first place. The timeline:

1. 1

   June 9: launch

   Fable 5 goes generally available as Anthropic's most capable model and the first Mythos-class release.
2. 2

   June 12: US export controls applied

   Three days in, US export controls are applied to the model, restricting availability outside the United States.
3. 3

   June 30: controls lifted

   The export controls are lifted after just under three weeks.
4. 4

   July 1: redeployed globally

   Anthropic redeploys Fable 5 worldwide with new cybersecurity classifiers, detailed at anthropic.com/news/redeploying-fable-5.

Fable 5 holds the best published SWE-bench scores of any generally available model, and Anthropic also cites it as #1 on Cognition's FrontierCode benchmark. The verified numbers:

Data verified July 18, 2026. Note what these numbers do and do not say: SWE-bench measures resolving real GitHub issues, which maps well to agentic coding, but no single benchmark settles a model war. OpenAI's GPT-5.6 Sol, for instance, leads on terminal-agent benchmarks, the full picture is in Claude Fable 5 vs GPT-5.6 Sol.

Fable 5 is not Claude Code's default model. You opt in explicitly, and you need Claude Code v2.1.170 or later:

/model fable

Two details worth knowing. The alias best resolves to Fable 5 if your organization has access, and to the latest Opus otherwise, so a script using it degrades gracefully. And Fable 5 is unavailable under zero-data-retention agreements, if your org runs ZDR, this model is off the menu regardless of budget.

If you run agents in parallel, the selection question gets more interesting. On an AIDEN board, each story card launches its own Claude Code session, so you can point Fable 5 at the one gnarly cross-cutting story while Opus and Sonnet handle the rest of the column. That per-story routing is the whole economics of the frontier tier, more in Claude Code orchestration.

Opus 4.8 is half the price and remains Anthropic's recommended starting point for complex agentic coding. Side by side:

The gap is real: Fable 5 is measurably ahead on the hardest benchmarks, and on long autonomous runs where a wrong turn costs hours, the stronger model pays for itself. Reach for it when the work is genuinely hard: architecture-level refactors that cut across a codebase, multi-hour unattended runs, and drafting specs against a large repo where the 1M tokens (128k output) window earns its keep.